Your documents hold the answer. They never leave your side.

Questions

What people ask

• What if the assistant gives a wrong answer?

  It will, sometimes — no retrieval system is perfect, and pretending otherwise would be the bigger risk. The acceptance test we write together before work starts sets the accuracy bar the pilot has to clear, on real questions against your real documents, not a demo set. The system is built to answer "that is not in the documents" rather than guess, and every answer carries its source passage, so your team verifies it in seconds instead of trusting it blindly.

• Our documents are confidential. What happens to them?

  They stay on your infrastructure — your server, or a dedicated EU machine registered to you. I work inside your controlled environment; your documents are never uploaded to my systems and nothing trains a model. Confidential is not the same thing as personal, and the distinction decides how the engagement is set up: the fixed pilot is scoped to corpora that carry no personal data — technical documentation, engineering manuals, standards and internal regulations, product catalogues, specifications. If the set you have in mind holds personal data instead, say so in your first message: that engagement is arranged and papered differently, and how it is classified is your DPO's call, not this page's.

• Why not just use ChatGPT (or a custom GPT) with our files?

  You can, for a quick internal experiment — and for some teams that is genuinely enough. I will say so rather than sell a pilot nobody needs. Where it breaks at scale: uploaded files sit on a third party's infrastructure, which is a hard no for regulated data; there is no acceptance test or measured accuracy; retrieval quality on large or messy document sets is unpredictable; and per-seat subscription cost scales with headcount, not with usage. The pilot exists to answer, with evidence, which side of that line your case is actually on.

• What does production actually cost after the pilot?

  The pilot itself is €4,500, fixed. Typical production rollouts — multi-source connectors, SSO, role-based access, more documents — run €10,000–20,000, and you get your real number only after the pilot proves retrieval works on your documents, not before. The optional Care Plan that follows starts from €1,500/month for monitoring, index updates and prompt tuning; nothing after the pilot is required.

• What happens if you disappear? You are one person.

  The system is built so the answer is "nothing dramatic": a standard open stack, full documentation, a runbook, and a handover pack your IT can pick up at any moment. That is a design requirement, not a courtesy — my own engineering rule is "will this still work when I can't be there?" The Care Plan exists for teams who prefer I keep maintaining it; leaving it is always possible, which is exactly why staying is safe.

• Do we need to buy or rent our own server for the pilot?

  Buying one is not required — renting a modest one is enough. A pilot-scale document set runs comfortably on a single dedicated EU server (Hetzner, OVH-class) rented in YOUR name for a few hundred euros a month — the same infrastructure story as the self-hosted AI track, just at pilot scale, not a fleet. Arranging that server is usually the first practical step, so it is worth starting before the three-week clock begins, not partway through it.